In order to improve the accuracy of flame recognition in different fire scenes,a flame identification and fire determination in complex scenes based on image processing was proposed.Combined with the application of Gaussian low-pass filtering(GLF),HSI-based block homogeneous filtering enhancement(HSI-BHF)and LAB-based K-means segmentation algorithm(LAB-Kmenas)to achieve flame recognition,it's determined whether a fire has occurred by calculating the area change rate and centroid dispersion of the flame.The results show that the GLF avoids the production of the "bell"phenomenon,and the PSNR are higher than the Butterworth and the ideal low-pass filtering. The HSI-BHF can suppress the influence of background brightness,achieve higher contrast compared with CLAHE and MSRCR.LAB-Kmeans reduces the effects of light and equipment,compared with Otsu and regional growth,it can more accurately extract the flame of in different scene fire images.Through the flame recognition test of 30 fire images in different scenes,the average recognition accuracy is 96.66%,and the average recognition time of one image is 1.94 s. Finally,taking the spreading fire of a restaurant and the steady flame of a candle as an example,after recognizing the video sequence image,it's determined that when the flame area change rate is not less than 0.22 and the centroid dispersion is not less than 17.02,the fire alarm should be triggered in time to reduce the loss of fire accidents.